SAW-Bench - Spatial Memory: leaderboard
Metric: Accuracy (%) on the 100 spatial memory questions (what changed in the scene between two visits) of the SAW-Bench multiple-choice questions about 786 egocentric videos recorded with Ray-Ban Meta (Gen 2) smart glasses (no audio); zero-shot; 2 frames per second where the model supports it (32 frames, or 8 for InternVL3 38B and InternVL2 40B, otherwise); answers parsed by regular expression, else by GPT-4o-mini; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 66 | #145 |
| 2 | Gemini 3 Flash (Preview) | 66 | #78 |
| 3 | GPT-5.2 (Medium) | 63 | #105 (GPT-5.2) |
| 4 | Gemini 3 Pro (Preview) (Low) | 63 | #64 (Gemini 3 Pro (Preview)) |
| 5 | GPT-5 Mini (2025-08-07) | 56 | #165 |
| 6 | InternVL3-14B | 54 | #494 |
| 7 | Qwen 3 VL 30B A3B (Thinking) | 54 | #338 (Qwen 3 VL 30B A3B) |
| 8 | Qwen 3 VL 32B (Thinking) | 52 | #287 (Qwen 3 VL 32B) |
| 9 | Gemini 2.5 Flash | 49 | #237 |
| 10 | Qwen 3 VL 8B (Thinking) | 48 | |
| 11 | InternVL3-8B | 46 | #606 |
| 12 | InternVL3-38B | 46 | #395 |
| 13 | Qwen 3 VL 235B A22B (Thinking) | 46 | #228 (Qwen 3 VL 235B A22B) |
| 14 | Qwen 2.5 VL 72B Instruct | 45 | #364 |
| 15 | Qwen 2.5 VL 32B Instruct | 45 | #443 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=saw-bench-spatial-memory · How It Works · Data refreshed daily, snapshot 2026-10-11.